Previously, the last push widget would only show when the branch never had
a merge request associated with it - even merged or closed ones. Now the
widget will disregard merge requests that are merged or closed.
Whenever you push to a branch GitLab will show a button to create a
merge request (should one not exist already). The underlying code to
display this data was quite inefficient. For example, it involved
multiple slow queries just to figure out what the most recent push event
was.
This commit changes the way this data is retrieved so it's much faster.
This is achieved by caching the ID of the last push event on every push,
which is then retrieved when loading certain pages. Database queries are
only executed if necessary and the cached data is removed automatically
once a merge request has been created, or 2 hours after being stored.
A trade-off of this approach is that we _only_ track the last event.
Previously if you were to push to branch A and B then create a merge
request for branch B we'd still show the widget for branch A. As of this
commit this is no longer the case, instead we will only show the widget
for the branch you pushed to most recently. Once a merge request exists
the widget is no longer displayed. Alternative solutions are either too
complex and/or too slow, hence the decision was made to settle for this
trade-off.
Performance Impact
------------------
In the best case scenario (= a user didn't push anything for more than 2
hours) we perform a single Redis GET per page. Should there be cached
data we will run a single (and lightweight) SQL query to get the
event data from the database. If a merge request already exists we will
run an additional DEL to remove the cache key.
The difference in response timings can vary a bit per project. On
GitLab.com the 99th percentile of time spent in User#recent_push hovers
between 100 milliseconds and 1 second, while the mean hovers around 50
milliseconds. With the changes in this MR the expected time spent in
User#recent_push is expected to be reduced down to just a few
milliseconds.
Fixes https://gitlab.com/gitlab-org/gitlab-ce/issues/35990
This commit migrates events data in such a way that push events are
stored much more efficiently. This is done by creating a shadow table
called "events_for_migration", and a table called "push_event_payloads"
which is used for storing push data of push events. The background
migration in this commit will copy events from the "events" table into
the "events_for_migration" table, push events in will also have a row
created in "push_event_payloads".
This approach allows us to reclaim space in the next release by simply
swapping the "events" and "events_for_migration" tables, then dropping
the old events (now "events_for_migration") table.
The new table structure is also optimised for storage space, and does
not include the unused "title" column nor the "data" column (since this
data is moved to "push_event_payloads").
== Newly Created Events
Newly created events are inserted into both "events" and
"events_for_migration", both using the exact same primary key value. The
table "push_event_payloads" in turn has a foreign key to the _shadow_
table. This removes the need for recreating and validating the foreign
key after swapping the tables. Since the shadow table also has a foreign
key to "projects.id" we also don't have to worry about orphaned rows.
This approach however does require some additional storage as we're
duplicating a portion of the events data for at least 1 release. The
exact amount is hard to estimate, but for GitLab.com this is expected to
be between 10 and 20 GB at most. The background migration in this commit
deliberately does _not_ update the "events" table as doing so would put
a lot of pressure on PostgreSQL's auto vacuuming system.
== Supporting Both Old And New Events
Application code has also been adjusted to support push events using
both the old and new data formats. This is done by creating a PushEvent
class which extends the regular Event class. Using Rails' Single Table
Inheritance system we can ensure the right class is used for the right
data, which in this case is based on the value of `events.action`. To
support displaying old and new data at the same time the PushEvent class
re-defines a few methods of the Event class, falling back to their
original implementations for push events in the old format.
Once all existing events have been migrated the various push event
related methods can be removed from the Event model, and the calls to
`super` can be removed from the methods in the PushEvent model.
The UI and event atom feed have also been slightly changed to better
handle this new setup, fortunately only a few changes were necessary to
make this work.
== API Changes
The API only displays push data of events in the new format. Supporting
both formats in the API is a bit more difficult compared to the UI.
Since the old push data was not really well documented (apart from one
example that used an incorrect "action" nmae) I decided that supporting
both was not worth the effort, especially since events will be migrated
in a few days _and_ new events are created in the correct format.
